Slide 1

CS 678 Boltzmann Machines1 Boltzmann Machine Relaxation net with visible and hidden units Learning algorithm Avoids local minima (and speeds up learning) by using simulated annealing with stochastic nodes

Slide 2

Node activation: Logistic Function Node k outputs s k = 1 with probability else 0, where T is the temperature parameter Node does asynchronous random update CS 678 Boltzmann Machines2

Slide 3

Network Energy and Simulated Annealing Energy is like in the Hopfield Network Simulated Annealing during relaxation Start with high temperature T (more randomness and large jumps) Progressively lower T while relaxing until equilibrium reached Escapes local minima and speeds up learning CS 678 Boltzmann Machines3

Slide 4

Boltzmann Learning Physical systems at thermal equilibrium obey the Boltzmann distribution P + (V ) = Probability that the visible nodes (V) are in state during training P - (V ) = Probability that the V is in state when free Goal: P - (V ) P + (V ) What are the probabilities for all states assuming the following training set (goal stable states)? 1 0 0 1 1 1 1 0 1 0 0 1 0 0 CS 678 Boltzmann Machines4

Slide 5

Boltzmann Learning Information Gain (G) is a measure of the similarity between P - (V ) and P + (V ) G = 0 if the probabilities are the same, else positive Thus we can derive a gradient descent algorithm for weight change by taking the partial derivative and setting it negative where p ij = probability that node i and node j simultaneously output 1 when in equilibrium CS 678 Boltzmann Machines5

Slide 6

Network Relaxation/Annealing A network time step is a period in which each node has updated approximately once 1. Initialize node activations (Input) Hidden nodes activations initialized randomly Visible nodes Random Subset of nodes set to initial state, others random Subset of nodes clamped, others set to random or initial state 2. Relax following an annealing schedule. For example: 2@30, 3@20, 3@10, 4@5 3. *Gather stats for m (e.g. 10) time steps, p ij = #times_both_on/m 4. Set final node state (output) to 1 if it was a 1 during the majority of the m time steps (could also output the probability or net value) CS 678 Boltzmann Machines6

Slide 7

Boltzmann Learning Algorithm Until Convergence (w < ) For each pattern in the training set Clamp pattern on all visible units Anneal several times calculating p + ij over m time steps end Average p + ij for all patterns Unclamp all visible units Anneal several times calculating p - ij over m time steps Update weights: w ij = C(p + ij - p - ij ) End CS 678 Boltzmann Machines7

Slide 8

4-2-4 Simple Encoder Example Map single input node to a single output node Requires log(n) hidden nodes 1. Anneal and gather p + ij for each pattern twice (10 time steps for gather). Noise.15 of 1 to 0,.05 of 0 to 1. Annealing Schedule: 2@20,2@15,2@12,4@10 2. Anneal and gather p - ij in free state an equal number of times 3. w ij = 2 (p + ij p - ij ) Average: 110 cycles CS 678 Boltzmann Machines8

Slide 9

4-2-4 Encoder weights before and after training Note common recursive weight representation What is the network topology? CS 678 Boltzmann Machines9

Slide 10

Shifting network, ~9000 cycles Note no explicit I/O directionality CS 678 Boltzmann Machines10

Slide 11

Boltzmann Learning But does this Boltzmann algorithm learn the XOR function Hidden nodes But first order weight updates (ala perceptron learning rule) CS 678 Boltzmann Machines11

Slide 12

Boltzmann Summary Stochastic Relaxation minima escape and learning speed Hidden nodes and a learning algorithm, improvement over Hopfield Slow learning algorithm but need to extend to learn higher order interactions A different way of thinking about learning creating a probabilistic environment to match goals Deep learning will use the Boltzmann machine (particularly the restricted Boltzmann machine) as a key component CS 678 Boltzmann Machines12